Unsupervised Sentence Representation via Contrastive Learning with Mixing Negatives
Authors: Yanzhao Zhang, Richong Zhang, Samuel Mensah, Xudong Liu, Yongyi Mao11730-11738
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The superior performance of the proposed approach is demonstrated via empirical studies on Semantic Textual Similarity datasets and Transfer task datasets. |
| Researcher Affiliation | Academia | 1Beijing Advanced Institution for Big Data and Brain Computing, Beihang University, Beijing, China 2SKLSDE, School of Computer Science and Engineering, Beihang University, Beijing, China 3Department of Computer Science, University of Sheffield, UK 4School of Electrical Engineering and Computer Science, University of Ottawa, Ottawa, Canada |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper states: "Our code is implemented in Python 3.6, using Pytorch 1.60 (Paszke et al. 2019), and the experiments are run on a single 32G NVIDIA A100 GPU." However, it does not explicitly state that the code is open-source or provide a link to it. |
| Open Datasets | Yes | For Semantic Textual Similarity (STS), we evaluate on seven datasets: STS12-16 (Agirre et al. 2012; Lee et al. 2013; Agirre et al. 2014, 2015, 2016), STS-B (Cer et al. 2017) and SICK-R (Marelli et al. 2014). For Transfer task (TR), we evaluate on seven datasets with the default configurations from Sent Eval.: MR (Pang and Lee 2005), CR (Kifer, Ben-David, and Gehrke 2004), SUBJ (Pang and Lee 2004), MPQA (Wiebe, Wilson, and Cardie 2005), SST-2 (Socher et al. 2013), TREC (Voorhees and Tice 2000) and MRPC (Dolan and Brockett 2005). |
| Dataset Splits | No | The paper states: "Following the standard evaluation protocol established in (Gao, Yao, and Chen 2021), we use the Sent Eval toolkit (Conneau and Kiela 2018) for evaluation purposes." and "We use the same training data and protocol in the work of Gao, Yao, and Chen (2021)." While these tools and references imply standard splits, the paper itself does not explicitly state the specific percentages or sample counts for training, validation, or test sets for its own experiments in the main text. |
| Hardware Specification | Yes | Our code is implemented in Python 3.6, using Pytorch 1.60 (Paszke et al. 2019), and the experiments are run on a single 32G NVIDIA A100 GPU. |
| Software Dependencies | Yes | Our code is implemented in Python 3.6, using Pytorch 1.60 (Paszke et al. 2019) |
| Experiment Setup | Yes | We set τ = 0.05, λ = 0.2 and use the Adam optimizer (Kingma and Ba 2014) for optimization. We experiment with the BERTbase and BERTlarge models using the respective learning rates 3e 5 and 1e 5. For both models, we train for one epoch with batch size 64. We use early stopping to avoid overfitting. |